You pick up your iPhone while waiting in line at a coffee shop. You google a not-so-famous actor, get linked to a Wikipedia entry listing his recent movies and popular YouTube clips of several of them. You check out user reviews on Amazon and pick one, download that movie on BitTorrent or stream that in Netflix. But suddenly the WiFi logo on your phone is gone and you're on 3G. Video quality starts to degrade, but you don't know if it's the server getting crowded or the Internet is congested somewhere. In any case, it costs you $10 per Gigabyte, and you decide to stop watching the movie, and instead multitask between sending tweets and calling your friend on Skype, while songs stream from iCloud to your phone. You're happy with the call quality, but get a little irritated when you see there're no new followers on Twitter. You may wonder how they all kind of work, and why sometimes they don't. Take a look at the list of 20 questions below. Each question is selected not just for its relevance to our daily lives, but also for the core concepts in the field of networking illustrated by its answers. This course is about formulating and answering these 20 questions. All the features of this course are available for free. It does not offer a certificate upon completion.
The Evolution of Online Advertising: Arel Lidow
Loading...
Из курса от партнера Princeton University
Networks: Friends, Money, and Bytes
62 оценки
Из урока
Guest Lectures
This contains various guest lectures from renowned members of academia and industry who are experts across the topics covered in this course.
Познакомьтесь с преподавателями
Mung ChiangProfessor
Electrical Engineering
Christopher BrintonLecturer
Electrical Engineering
Alright so big [inaudible] my name is Earl Lido.
I actually took this class in 2004, I guess the predecessor to this class.
I think, there were like five people in the class.
So, it's great to see that it's grown over time, It's really a pleasure to come back
here every year and talk to you guys, I actually worked in finance before this,
and spoke about what I did then. And then, have since moved on to online
advertising. So there are a lot of cool problems, in
online advertising. And today, I just wanna back up and sort
of tell you guys the, the story of how all this evolved.
It's really easy to look at, you know, a textbook and see a, a really elegant
theory that's been laid down. But, but the truth is most of this most of
how online advertising came to use auctions was actually very organic.
Sort of evolved naturally over time So I think to understand it you really have to
flash back to 1994. I realized for the first time that many of
you in this class may not actually remember 1994.
But as hard as it is to believe, in 1994 people really didn't know what the
internet was. It was just starting to enter the public
consciousness. When people thought of online, they really
thought of America Online. Like you'd sign on, you'd check your
email. There are chat rooms.
It was really cool back in the day. And people were just starting to hear
about, about the world wide web. Companies were just starting to build
websites. And most importantly, 1984 was the year
that online advertising was invented. So this was the first online ad ever
shown, actually. It was probably the most deceptive ad ever
shown. It literally says, I can't read it.
Have you ever clicked your mouse here? [laugh].
And it was it was an AT&T ad, as hard as it is to believe.
And it did pretty well. Like, like, most people had no idea what
this was. Seven out of ten people clicked on it.
Which, by today's standards, is really, really good.
It's, you're really lucky today if you get one out of every 1,000 people to click on
an ad. So this is a very successful ad.
And this really was the birth of a new industry.
It's now, you know. $40 billion a year industry.
It's hard to believe. Back, back then in early 90's, You know, how you got an ad on
a website was a very simplistic model. You literally picked up a telephone and
you called up the website, or sales team at a website, and you said, I want to put
my ad here. And you haggle a little bit over price and
you basically arrive at some CPM, a cost per thousand impression value, that you're
going to pay. And your ad would go up for a certain
amount of time. There wasn't a lot of sophistication to
it. You know It'll be really interesting to
see the equations for this because, this is very simple just haggle over price and
decide. So not much actually changed until sort of
late 90s. I think this was the first revolution in
online advertising. This was the year that Britney Spears
became famous. I don't know if you guys remember that.
Do you guys remember. Like Britney Spears came out with, Baby
One More Time? It was also the height of the dot.com
bubble. Pets.com somehow raised 83 million dollars
without a business model. It was amazing.
And it was also the year that, Idea Lab, which was, like, the Y combinator of its
day launched a company called goto.com. Here's the homepage of goto.com.
It actually looks pretty similar, you know, despite the ugly logo here.
It looks pretty similar to Google and their, their model there business model is
actually very similar as well. They set out to disrupt the, the
classifieds industry. So their, their business model was that
they built a search engine. Actually, they licensed a search
technology from Intomi, that allowed people to search for whatever they were
looking for. And they'd allow advertisers to pay to
have their advertisements actually interspersed with the actual search
results. It'd be like, you know, getting Google
search except the ad, you can't tell what's an ad, and what's an actual search
result. And this model actually took off.
Like, for a while people didn't really know whether GoTo would be the big search
engine of the future or Google. So it's actually a fairly successful
business, and here's how they sold their ad spots.
So they, they built a self service platform, and advertisers could log in,
and they could put in a bid value, an amount that they were willing to pay per
click. And whoever bid the highest per click got
placed up higher in the search results. Now, I think somebody over here asked a
really good question about why that actually matters.
Like, why is it better to be placed up higher?
Look here it matters a lot you know, if you're, cause people are generally assumes
unconsciously that whatever comes up first is more relevant, most relevant search
result. So, you couldn't tell what was a search
result and what was mad. Then whoever was at the very top got all
the clicks. But their model was very simplistic.
It was not a second price option. Whoever bid the highest per click got
highest in the search results. This had some serious issues with it.
So the biggest issue was that buyers figured out that they could game the
system. So they log in, and people figure out what
actually happens if I lower my bid, Could I get the same placement?
And you know, the top buyer actually values you know, a click at $five.
In this case, they can actually shave their bid, they can lower their bid to
$four and a penny, and they can still get the top spot.
So what happened was, literally, buyers would log in and they would change their
bid thousands of times per day. This wreaked havoc on GoTo's
infrastructure. It was very, very, like, difficult for
them to maintain this auction system. And as a result, they also may have lost
money. Over time everybody stayed at their bid,
there's a lot of variation in that, but over time prices were actually lower than
on Google, which we'll talk about. So, fast forward to 2000, 2001.
So, a new company came along called Google.
Which, you know I see a Google shirt over here.
Okay. So, I guess people have heard of it.
And, I think it was actually 2001 that they launched their self serve AdWords
products. And this really changed everything.
You're not only did Google have the best search engine out there, but they really
innovated the way that ad space was sold. So they had two major innovations and both
of these were completely organic. They,
They weren't based on theory. Nobody read a textbook.
According to In The Plex, like the book about Google was actually just the
engineer working on this had these two absolutely brilliant ideas.
So first was he implemented the option as second price option, generalize second
price option. And as you guys know, what this means is
whoever bids highest actually pays what the second highest bidder bid.
So in this case if the top. Buyer.
If the top bidder bids $five per click. And the second highest bidder bids $three
per click, then the top bidder wins the top spot, but only pays $three.
So this incentivizes, by, incentivizes advertisers to actually bid their fair
value. So, in this case, if the top bidder bids
$four they are still going to pay $three, if they bid at $five, they still pay
$three. So they might as well bid their fair
value. You don't see the same effect of what
happened at goto.com where everybody would play around with their bids constantly and
eventually it resulted in much lower bids. The second major innovation which actually
seems very simple in retrospect. It's hard to believe that nobody else
think about this. Was it google tied valuation to click
through rate. What this means is that they, you know of
course you don't realized that just because somebody bids more per click
doesn't necessarily mean that Google is going to make more revenue or effectively
more profit off of that. So if the top bidder here is willing to
pay $five per click. And the second guy is willing to pay
$three per click. But the top buyer get, top bidder gets one
click every hour and the second bidder gets two clicks every hour.
Google's going to make $six by putting the, the second highest bidder's bid up
top. First it was $five.
So this results in a very different ranking of, of how, how ads are actually
shown on a page. And as a result, it maximizes revenue for
the seller you know Google made a lot more money this way.
It's actually kind of shocking to believe. So goto.com was later renamed as Overture,
and sold to Yahoo for about $two billion. And Yahoo, for as long as they were
running their own search engine, actually used a first, basically a first price
auction model and didn't factor in actual click through rate into their, to their
evaluation. So a lot has been, a lot academic research
has been done on Google's auction model and people have, you know, rightfully
discovered that it has a few inefficiencies net.
So one of the, the biggest things and, by the way, it's worth noting that Google
doesn't actually use a pure generalized second price auction anymore.
There's a lot more that goes into it. So this is a bit of an over
simplification. One of the, the big issues with
generalized second price auctions in general is what happens if a bidder values
two of the same goods or two ad spots, in this case, at effectively the same value.
So, what if I'm a bidder and I realize that I get about the same click through
rate. If I'm in the first placement or the
second placement. So in this case I got one click every 60 minutes, if I'm not at
the very top and they get about the same one click every 61 minutes if I'm in the
second highest position. So in this case, if, if I value a click at
$five and the second highest bidder values a click at $3.00, why would I pay $five to
get a click? Why wouldn't I pay2.99 $2.99 just below
the second bidder? In that case I'm going to pay a heck of a
lot less for something that's worth just about the same amount as me.
Because as an advertiser I really care about how often people are clicking on my
ads. So this creates an issue.
There's a lot of, you know, research into how to actually solve this problem.
But, as you just learned Victory Clarke Groves or VCG is sort of the economically
right way to solve this. I'm not going to go into a lot of depth
about it. It actually gets kind of complicated.
And the basic idea is this. So we have two bidders and we're
auctioning off, essentially, two goods. So this is why I'm not in academia.
I think in terms of apples and oranges, [laugh] not equations.
But, but here is a very simple example. We have an apple and orange.
We're auctioning off both the apple and the orange.
We have two bidders. So bidder A says, I'll pay $ten for the
apple but $nine for the orange. I've valued them, you know, roughly the
same. Bidder B says I'll pay $nine for the apple
and $eight for the orange. So in this case, you know, bidder A could
just. Like bid $seven.
Get roughly equivalent value for, for the orange but here's how to solve the
problem. So in Victory Clarke Groves, VCG the
bidder actually pay, the winning bidder actually pays the amount of harm they
cause to the second highest bidder. So in this case, if bidder A wins the
apple, say bid $ten they value at $ten they effectively cause a dollar of harm to
the second bidder. The second bidder now has to live with
this orange. They're are willing to value the apple.
They, they would value the apple at $nine. So they are effectively living with a $one
less of value that they could've had. So by, by charging bidder A just $one.
There's no incentive for, for bidder A to actually bid less purposely to get a worse
bid, but get something that is roughly equivalent in value.
So again this, this actually gets kind of complicated.
Won't go into it much. In practice it's worth noting that not too
many people actually use VCG. The reason being is that sellers actually
make less money and it's also very difficult to explain to buyers.
Especially if you're running a self-service ad platform explaining
something like VCG, which is very complicated, is just too difficult.
So Google makes more money by not using VCG and it's easier to explain to buyers.
It's also more computationally difficult to, to implement.
So fast forward to current day. So Google has obviously been very
successful with that model. They, they're on track to make about $40
billion in revenue off of advertising. Facebook is basically the new AOL.
They, you know, a third of all page views on the Internet are through Facebook now.
But shockingly most Internet ads are actually shown in a very inefficient way.
Remember that first slide I showed, where somebody picks up the phone and calls up a
website? That's effectively what happens for about
70% of all the money spent on banner advertising.
It's banners, you see literally somebody at agency picks up a phone and calls up an
ad network or a website and says, you know I'll pay you this much to show you my ad.
Gets a little bit more sophisticated, but, but that's basically the general jist of
it. It's absolutely shocking that this is the
case, You know, 2012, we have all this really
sophisticated technology, and yet most, most online advertising is still done by
somebody haggling over a phone. So App Nexus was founded, 2007, based in
New York City. And, we're basically trying to solve this
problem. So what we do is, we, we replace that
phone with an auction mechanism. We feel that this is more efficient at, at
serving the right ad to the right user. What we do is, is, we, every time a user
goes to the website, that site makes a request to our servers, and it says, Aoo
Nexus go run an auction and figure out which ad we should show here.
Then we reach out, We make a request to thousands of
different advertisers all within about ten milliseconds.
And we say, you know Amazon, how much are you willing to pay based on what you know
about this user? Ebay, how much are you willing to pay?
We get bids back from all these different bidders.
And keep in mind, they're all bidding based on what they know about the user.
So if you recently bought something from eBay, eBay might say, oop, this user has a
higher propensity to buy so we're going to bid more on them.
So everybody returns their bids, We select the highest bidder.
So, this basically replaces the phone. It's like real time programmatic haggling
essentially. And you know just like Google and AdWords,
we used a second price auction as well. So, very simple example.
Ebay is bidding two cents, GroupOn bids a penny.
Ebay would win. They get to show their ad, but they would
only pay a penny. Now, in some sense, this is a much simpler
problem than Google has we're not auctioning off multiple ad spots on a
page. We're auctioning off essentially, one at a
time. But we do this 30 billion times a day.
So, we run a lot of auctions [laugh]. But there's a lot of hidden complexity to
this specifically, I mean, we know some major inefficiencies with a second price
auction. So, here's, here's a good example of one.
So, so in banner advertising how many times and what frequency you share your
ads to, to user? Makes a lot of difference for an
advertiser. So if I'm Ford and I show you an ad once,
you know, that ad is probably going to be somewhat effective.
If I showed it to you a second time, you've already seen the ad so it's going
to be a little bit less effective, especially if I show it to you one right
after the other. If I show you my ad 100 times, you're
probably going to get annoyed. That might actually have negative value to
me. So naturally, bidders tend to bid less
after they've shown a user their ad. So the first time, you know, Ford sees a
user they might bid $five. The second time they'll bid $2.50.
Third time might bid $1.25 and so on. They're going to keep decreasing their bid
usually some exponential function over time.
So I modeled this act here. Here's a very simple example of this happening.
So, here we have Ford, AT&T, and Kraft. And this is the number of times they have
seen the same user, over and over again. So the first time they all see a user Ford
bids $five. At&T bids $three, and Kraft bids $two, And
so Ford is going to win that auction. They bid the most and they are going to pay the
second price which is what AT&T bid $three.
And the second time everybody sees, sees that user so now Ford is bidding less.
They are only bidding $2.50 because they've already shown that to the user, to
the user so the user is less valuable to them.
At&t now has the highest bid. At&t is still bidding $three cuz they
haven't had a chance to show that ad to the user yet.
And Kraft bid $two So, AT&T is going to win that auction, and it's going to pay
$2.50, the second highest price, and so on and so forth.
So eventually, you know, the price naturally goes down the more frequently
you see a user. Because everybody's had a chance to show
their ads to that user. So, in this case, you know, the average
bid price is $1.29. And the seller is making about $thirteen
in total. Now, what happens if Ford, you know the
highest bidder in this case says you know, we're gonna try something a little bit
different. For the first five times we see this user,
we're not going to submit a bid. We're not going to bid at all and the
sixth time we see the user, we're going to start bidding then.
But we're actually going to start bidding much less. So, what, what happens?
So in this case, we're still running a second price auction.
But here, you know, Ford isn't bidding for the first five auctions.
In the first auction, AT&T bids $three, Kraft bids $two.
Kraft ends up paying $two right? Less than what, what they would have made
in the previous case. And so on and so forth, until the sixth
time that we see that user. This time, Ford bids, and they bid $three
but because AT&T and Kraft, the other two advertisers have already shown users their
ads so many times, They're now bidding at 33 cents.
So Ford wins the option as they bid $three but way less than they would have paid
before but they're only paying 33 cents now whereas they paid $three before paying
a tenth of that now. It's kind of crazy.
So now all of a sudden the average price, the average bid price goes down to 33
cents and the seller only makes $seven almost half of what they made before.
Yeah. Doesn't this strategy in which they bid
over and over and over again hurt Ford? And what you said, the annoying of the
user if, I don't know. If you have the, the ad show up multiple
times? So, so you're, so you're saying, isn't it,
so, it's less valuable the second time you show an ad to a user.
Yeah. Yeah.
So the answer is buyers will, advertisers will typically bid to show the user an ad
a second time. They'll just bid less, because, as you
say, it's less valuable. Yeah.
In some cases there's a lot of very sophisticated stuff that goes on, so they
may not bid at all. They may say, it's just not worth it for
us to show an ad the second time or we're going to show our ads to the user every
other time, something like that. This is a very simple example of that.
Yeah? Isn't this from, [inaudible], that you
have a [inaudible]? Yes, so as your bid density increases you
have this problem much less. But also depends on, so when, when you
take into account what people know, what different bidders know about the user,
then it becomes very different because on any given option you may have, let's say,
you know Ford, AT&T, and Kraft and the 100 other bidders may be participating but
Ford, AT and T, Kraft may be the only advertisers that know something about that
user. So Ford may know that you've been to their
website recently, you know Kraft may know you've recently bought their Oreos or
something. At and T may know that you're currently a
Verizon subscriber. Because they now these things they're all
bidding on a separate good. They're, they're actually bidding on much
more than everybody else. So all of a sudden you have a situation
where you basically, you know have a much lower bid density at a much higher price
point. This actually happens quite a bit.
Yeah? As a bidder, how am I choosing which sites
I would want my banners to go onto? Like which sites [inaudible]?
Yeah, so that's an excellent question. So there's actually, so, there's a lot of
thought that goes into this. Typically, you, you serve a, an
optimization process. So, over time, you basically run your ads
everywhere, or where you think that they're going to work.
And then you look at what sites are actually working.
To where you get the most clicks, or where people sign up for AT&T subscriptions.
And over time you start to spend your money in a more, like more concentrated
way just on those sites. It's a very simplistic way of doing it but
there is, it actually gets pretty complex. How you do at opposition.
Sorry, you had a question before. I don't think it would make any difference
in this option, but why did you lower Ford's first bid price?
Would it make any difference to [inaudible].
Yeah, well I mean just to demonstrate they could actual bid less than their fair
value and still win. Why would they want to do that? Why would
they want to in this case? It actual doesn't make a difference in
this case but I'm just pointing out that there's no, there's nothing preventing
them from not telling the truth in this case.
Yeah. If you have a small number of bidders,
couldn't the bidders just like, collaborate and say, you're going to bid
two cents, I'm going to bid one cent, and we're going to take turns?
Yeah, collision its a big problem. [laugh].
[laugh]. We have like, its the result of.
Like, [inaudible] way to get around that, or do you just have to, like, have
[inaudible]? I mean, you, you can, you can look for
patterns. So you can actually, you know, look to see
what the average price of an ad placement is, for instance.
And who's bidding on it and if you see sudden drops in the price or, you know,
the price tends, trends down over time. Then you can kinda figure out what's going
on. It's actually very difficult to, to figure
out in practice. And it does happen.
There, there are all sort of other tricks that, that bidders use.
Say another one is called bid stuffing. So they'll, if, so each of these buyers
probably has an, a budget. An amount of money that they're willing to
spend to show ads. So, in this case, if AT&T knows that
Ford's budget is $100, let's say. They could bid purposely higher to get
Ford to spend more money. And, to spend out their budget faster so
then, all of a sudden, AT&T can start winning at a lower price.
Yeah. [inaudible] another answer [inaudible].
You can have like a minimum. The separation between [inaudible],
continuity. If they did know each other, they can't
influence the new web site. Yeah, right, so that, that's one, one
methodology for doing that. There's, I read some papers on like,
random enterprise auctions, so there's some,
Some theory, I, I don't think it's actually been tried very well in practice.
But what, what would happen if you actually ran a third price auction or a
fourth price auction? And what happens if the buyers actually
don't know which level they're being price reduced down to?
Then it significantly changes that dynamics, and actually this incentivizes
things like collision. Yeah.
In this case, does each buyer know how much the other people are bidding? in this
case they don't. Yep.
What about a system where once people drop out of the auction, you're out of the
auction for the entire cycle? Once people drop out of it?
Yeah, so like. So, so you can never bid on the user
again. Yeah, not that.
That, that's. You can never bid on that user, but maybe
for this specific time bracket. Like, can't bid for the next five minutes
or. Yeah, exactly.
Hour or something. Once you drop from the auction.
It's an interesting idea, interesting idea.
I don't know. I mean that, that would potentially be
damaging to sellers because it decreases bid density on subsequent bids.
But, but it's an interesting idea, and so we're trying.
So, so here's what sellers typically do to prevent against this type of practice.
So, so the best defense that they have today.
Is called price floors. This is the equivalent of, of using a
reserve price on Ebay. So sellers generally know, you know, what
their ad placements are worth to buyers. They, they know roughly.
So what they do is they say, you know we're not going to allow buyers to buy our
ad placement for under this dollar value. So, in this case taking the exact same,
exact same example as before but simulate it with $1.25 cent price floor.
So what happens? So in this case, we run the first five
auctions. In the AT&T and Kraft bid we notice that
on auctions four and five and the seller actually chooses not to show an ad at all.
So they don't accept anybody's bid. Cuz they, they know that they're probably
not getting a fair value for, for their inventory.
On the sixth option, you know, Ford comes in and they're going to get price reduced,
but they're actually going to get price reduced down to the value of that price
floor. So seller says, you know, I'll sell, I'll
sell my ad placement, but for no less than $1.25.
So this case, we're not price reducing down to the second bid, we're price
reducing down to the value of the price floor.
So Ford actually pays $1.25. On the seventh bid, they're also going to
pay $1.25. So here the, the seller actually ends up
making a little bit more money than they made before.
So, seller made about $seven before. Now they're making $7.25.
So seller's actually increased their revenue by putting in these price floors
and actually not showing bid, not showing ads on some bids.
But this also gets to be a very risky game for sellers.
Because what happens if you set a price floor too high or too low?
So here's a, the same example, but with $1.50 price floor on the right hand side.
So see what happens here, is that the seller chooses not to show an ad more
often, that, you know, Ford is gonna pay more when they actually win.
But the seller actually makes less money than they would have made if they had no
price wars. So a thing at $seven before, now they're
only making about $6.50. So, this ends up being a pretty
complicated game that buyers and sellers play.
This actually happens in real life, like, it's not just theory, it actually happens.
So here's an example of some real data. I can't tell you which website this is,
but, but, but I promise you these are two different websites here.
On the top, so on the Y axis on both these charts, is the win rate.
So this is how often buyers are winning as a function of how much they're bidding.
So on the, the top chart here, we see that it's, you know, roughly linear.
It's basically what we'd expect. If buyers bid a higher price, they'll win
more often. But on the slower chart here you see some
interesting behavior. So.
Buyers never win if they bid less than 90 cents, meaning the se, the seller here has
probably put in a price floor at 90 cents. So nobody can, nobody can win my ad
placement if they pay less than 90 cents. What happens is, okay, well, after 90
cents, then everybody's bidding just around 90 cents and is winning.
So in, in this case, the seller's probably put in a price floor that's actually too,
too high. Because everybody's winning after 90
cents, so there's no distri, there's not a linear distribution of bids after that 90
cent price floor. So this ends up being a pretty complicated
game that we dealt with everyday at App Nexus.
Ознакомьтесь с нашим каталогом
Присоединяйтесь бесплатно и получайте персонализированные рекомендации, обновления и предложения.
Coursera делает лучшее в мире образование доступным каждому, предлагая онлайн-курсы от ведущих университетов и организаций.
© Coursera Inc., 2019 Все права защищены.
